"""Template for showing the results of the last experiment in MLFlow."""

import logging
import numpy as np
import helpsk as hlp
import pandas as pd
import plotly_express as px
from helpsk.utility import read_pickle, Timer
from helpsk.sklearn_eval import MLExperimentResults

from source.service.model_registry import ModelRegistry

%cd /code

from source.config import config  # noqa: E402
logging.config.fileConfig(
    "source/config/logging_to_file.conf",
    defaults={'logfilename': 'output/log.log'},
    disable_existing_loggers=False,
)

/usr/local/lib/python3.11/site-packages/IPython/core/magics/osm.py:417: UserWarning: using dhist requires you to install the `pickleshare` library.
  self.shell.db['dhist'] = compress_dhist(dhist)[-100:]

/code

Get Latest Experiment Run from MLFlow

registry = ModelRegistry(tracking_uri=config.experiment_server_url())
experiment = registry.get_experiment_by_name(exp_name=config.experiment_name())
logging.info(f"Experiment id: {experiment.last_run.exp_id}")
logging.info(f"Experiment name: {experiment.last_run.exp_name}")
logging.info(f"Run id: {experiment.last_run.run_id}")
logging.info(f"Metric(s): {experiment.last_run.metrics}")

2023-11-24 21:00:13 - INFO     | Experiment id: 1
2023-11-24 21:00:13 - INFO     | Experiment name: credit
2023-11-24 21:00:13 - INFO     | Run id: 7244e48d74144410aa67215f5b59ef38
2023-11-24 21:00:13 - INFO     | Metric(s): {'roc_auc': 0.762767489489011}

Last Run vs Production

What is the metric/performance from the model associated with the last run?

logging.info(f"last run metrics: {experiment.last_run.metrics}")

2023-11-24 21:00:13 - INFO     | last run metrics: {'roc_auc': 0.762767489489011}

What is the metric/performance of the model in production?

production_run = registry.get_production_run(model_name=config.model_name())
logging.info(f"production run metrics: {production_run.metrics}")

2023-11-24 21:00:13 - INFO     | production run metrics: {'roc_auc': 0.753377535324465}

Last Run

# underlying mlflow object
experiment.last_run.mlflow_entity

<Run: data=<RunData: metrics={'roc_auc': 0.762767489489011}, params={'model__criterion': 'entropy',
 'model__max_depth': '70',
 'model__max_features': '0.1142268477118407',
 'model__max_samples': '0.5483119512487002',
 'model__min_samples_leaf': '8',
 'model__min_samples_split': '12',
 'model__n_estimators': '553',
 'prep__numeric__imputer__transformer': 'SimpleImputer()',
 'prep__numeric__pca__transformer': "PCA(n_components='mle')",
 'prep__numeric__scaler__transformer': 'None',
 'prep__savings_status__savings_encoder__transformer': 'SavingsStatusEncoder()'}, tags={'mlflow.log-model.history': '[{"run_id": "7244e48d74144410aa67215f5b59ef38", '
                             '"artifact_path": "model", "utc_time_created": '
                             '"2023-11-24 21:00:08.619307", "flavors": '
                             '{"python_function": {"model_path": "model.pkl", '
                             '"predict_fn": "predict", "loader_module": '
                             '"mlflow.sklearn", "python_version": "3.11.6", '
                             '"env": {"conda": "conda.yaml", "virtualenv": '
                             '"python_env.yaml"}}, "sklearn": '
                             '{"pickled_model": "model.pkl", '
                             '"sklearn_version": "1.3.2", '
                             '"serialization_format": "cloudpickle", "code": '
                             'null}}, "model_uuid": '
                             '"8e654c1d76154dbba04829c84f68f68a", '
                             '"mlflow_version": "2.8.0", "model_size_bytes": '
                             '2595199}]',
 'mlflow.note.content': '2023_11_24_20_59_43',
 'mlflow.runName': '2023_11_24_20_59_43',
 'mlflow.source.git.commit': '81a963fcbc4794b8b7bc6c330fc6b034760eb65d',
 'mlflow.source.name': 'source/entrypoints/cli.py',
 'mlflow.source.type': 'LOCAL',
 'mlflow.user': 'root',
 'type': 'BayesSearchCV'}>, info=<RunInfo: artifact_uri='/code/mlflow-artifact-root/1/7244e48d74144410aa67215f5b59ef38/artifacts', end_time=1700859610232, experiment_id='1', lifecycle_stage='active', run_id='7244e48d74144410aa67215f5b59ef38', run_name='2023_11_24_20_59_43', run_uuid='7244e48d74144410aa67215f5b59ef38', start_time=1700859583332, status='FINISHED', user_id='root'>, inputs=<RunInputs: dataset_inputs=[]>>

---

Load Training & Test Data Info

with Timer("Loading training/test datasets"):
    X_train = experiment.last_run.download_artifact(artifact_name='x_train.pkl', read_from=read_pickle)  # noqa
    X_test = experiment.last_run.download_artifact(artifact_name='x_test.pkl', read_from=read_pickle)  # noqa
    y_train = experiment.last_run.download_artifact(artifact_name='y_train.pkl', read_from=read_pickle)  # noqa
    y_test = experiment.last_run.download_artifact(artifact_name='y_test.pkl', read_from=read_pickle)  # noqa

Timer Started: Loading training/test datasets
Timer Finished (0.01 seconds)

logging.info(f"training X shape: {X_train.shape}")
logging.info(f"training y length: {len(y_train)}")

logging.info(f"test X shape: {X_test.shape}")
logging.info(f"test y length: {len(y_test)}")

2023-11-24 21:00:13 - INFO     | training X shape: (800, 20)
2023-11-24 21:00:13 - INFO     | training y length: 800
2023-11-24 21:00:13 - INFO     | test X shape: (200, 20)
2023-11-24 21:00:13 - INFO     | test y length: 200

np.unique(y_train, return_counts=True)

(array([0, 1]), array([559, 241]))

train_y_proportion = np.unique(y_train, return_counts=True)[1] \
    / np.sum(np.unique(y_train, return_counts=True)[1])
logging.info(f"balance of y in training: {train_y_proportion}")

2023-11-24 21:00:13 - INFO     | balance of y in training: [0.69875 0.30125]

test_y_proportion = np.unique(y_test, return_counts=True)[1] \
    / np.sum(np.unique(y_test, return_counts=True)[1])
logging.info(f"balance of y in test: {test_y_proportion}")

2023-11-24 21:00:13 - INFO     | balance of y in test: [0.705 0.295]

Cross Validation Results

Best Scores/Params

results = experiment.last_run.download_artifact(
    artifact_name='experiment.yaml',
    read_from=MLExperimentResults.from_yaml_file,
)
logging.info(f"Best Score: {results.best_score}")
logging.info(f"Best Params: {results.best_params}")

2023-11-24 21:00:13 - INFO     | Best Score: 0.762767489489011
2023-11-24 21:00:13 - INFO     | Best Params: {'model': 'RandomForestClassifier()', 'max_features': 0.1142268477118407, 'max_depth': 70, 'n_estimators': 553, 'min_samples_split': 12, 'min_samples_leaf': 8, 'max_samples': 0.5483119512487002, 'criterion': 'entropy', 'imputer': 'SimpleImputer()', 'scaler': 'None', 'pca': "PCA('mle')", 'savings_status_encoder': 'SavingsStatusEncoder()'}

# Best model from each model-type.
data = results.to_formatted_dataframe(return_style=False, include_rank=True)
data["model_rank"] = data.groupby("model")["roc_auc Mean"].rank(method="first", ascending=False)
data.query('model_rank == 1')

	rank	roc_auc Mean	roc_auc 95CI.LO	roc_auc 95CI.HI	model	C	max_features	max_depth	n_estimators	min_samples_split	...	colsample_bytree	colsample_bylevel	reg_alpha	reg_lambda	num_leaves	imputer	scaler	pca	savings_status_encoder	model_rank
11	1	0.763	0.726	0.800	RandomForestClassifier()	NaN	0.114227	70.0	553.0	12.0	...	NaN	NaN	NaN	NaN	NaN	SimpleImputer()	None	PCA('mle')	SavingsStatusEncoder()	1.0
0	2	0.759	0.713	0.805	LogisticRegression()	NaN	NaN	NaN	NaN	NaN	...	NaN	NaN	NaN	NaN	NaN	SimpleImputer()	StandardScaler()	None	OneHotEncoder()	1.0
19	3	0.758	0.717	0.800	XGBClassifier()	NaN	NaN	1.0	896.0	NaN	...	0.906088	0.825022	0.003169	1.410800	NaN	SimpleImputer(strategy='median')	None	None	SavingsStatusEncoder()	1.0
24	5	0.755	0.723	0.787	LGBMClassifier()	NaN	NaN	NaN	NaN	NaN	...	0.682707	NaN	5.683878	42.573842	50.0	SimpleImputer(strategy='median')	None	PCA('mle')	OneHotEncoder()	1.0
5	9	0.747	0.698	0.796	ExtraTreesClassifier()	NaN	NaN	NaN	NaN	NaN	...	NaN	NaN	NaN	NaN	NaN	SimpleImputer()	None	None	OneHotEncoder()	1.0

5 rows × 26 columns

results.to_formatted_dataframe(return_style=True,
                               include_rank=True,
                               num_rows=500)

rank	roc_auc Mean	roc_auc 95CI.LO	roc_auc 95CI.HI	model	C	max_features	max_depth	n_estimators	min_samples_split	min_samples_leaf	max_samples	criterion	learning_rate	min_child_weight	subsample	colsample_bytree	colsample_bylevel	reg_alpha	reg_lambda	num_leaves	imputer	scaler	pca	savings_status_encoder
1	0.763	0.726	0.800	RandomForestClassifier()	<NA>	0.114	70.000	553.000	12.000	8.000	0.548	entropy	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer()	None	PCA('mle')	SavingsStatusEncoder()
2	0.759	0.713	0.805	LogisticRegression()	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer()	StandardScaler()	None	OneHotEncoder()
3	0.758	0.717	0.800	XGBClassifier()	<NA>	<NA>	1.000	896.000	<NA>	<NA>	<NA>	<NA>	0.029	8.000	0.799	0.906	0.825	0.003	1.411	<NA>	SimpleImputer(strategy='median')	None	None	SavingsStatusEncoder()
4	0.758	0.708	0.807	RandomForestClassifier()	<NA>	0.681	38.000	1,461.000	23.000	10.000	0.553	gini	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer(strategy='median')	None	None	OneHotEncoder()
5	0.755	0.723	0.787	LGBMClassifier()	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	0.511	0.683	<NA>	5.684	42.574	50.000	SimpleImputer(strategy='median')	None	PCA('mle')	OneHotEncoder()
6	0.755	0.720	0.789	LGBMClassifier()	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	0.797	0.700	<NA>	6.654	9.475	381.000	SimpleImputer(strategy='median')	None	PCA('mle')	OneHotEncoder()
7	0.752	0.711	0.792	RandomForestClassifier()	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer()	None	None	OneHotEncoder()
8	0.751	0.682	0.820	LogisticRegression()	0.000	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer(strategy='most_frequent')	MinMaxScaler()	None	OneHotEncoder()
9	0.747	0.698	0.796	ExtraTreesClassifier()	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer()	None	None	OneHotEncoder()
10	0.741	0.691	0.791	RandomForestClassifier()	<NA>	0.710	15.000	1,493.000	33.000	27.000	0.914	gini	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer()	None	PCA('mle')	OneHotEncoder()
11	0.739	0.671	0.807	LogisticRegression()	23.327	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer()	StandardScaler()	None	OneHotEncoder()
12	0.738	0.705	0.772	LGBMClassifier()	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer()	None	None	OneHotEncoder()
13	0.737	0.667	0.808	ExtraTreesClassifier()	<NA>	0.030	84.000	1,088.000	24.000	36.000	0.981	gini	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer()	None	None	SavingsStatusEncoder()
14	0.735	0.683	0.787	LGBMClassifier()	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	0.598	0.661	<NA>	12.533	35.084	348.000	SimpleImputer(strategy='most_frequent')	None	PCA('mle')	SavingsStatusEncoder()
15	0.733	0.666	0.800	LogisticRegression()	0.000	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer(strategy='most_frequent')	StandardScaler()	None	OneHotEncoder()
16	0.732	0.666	0.798	ExtraTreesClassifier()	<NA>	0.857	30.000	879.000	17.000	28.000	0.563	entropy	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer(strategy='median')	None	None	SavingsStatusEncoder()
17	0.732	0.664	0.799	ExtraTreesClassifier()	<NA>	0.672	81.000	1,136.000	34.000	34.000	0.971	gini	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer(strategy='median')	None	None	OneHotEncoder()
18	0.731	0.681	0.782	XGBClassifier()	<NA>	<NA>	5.000	1,218.000	<NA>	<NA>	<NA>	<NA>	0.115	2.000	0.545	0.648	0.852	0.123	1.165	<NA>	SimpleImputer(strategy='median')	None	PCA('mle')	OneHotEncoder()
19	0.729	0.668	0.791	LGBMClassifier()	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	0.845	0.453	<NA>	16.166	40.978	351.000	SimpleImputer(strategy='median')	None	PCA('mle')	SavingsStatusEncoder()
20	0.729	0.674	0.783	LogisticRegression()	0.001	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer(strategy='median')	StandardScaler()	PCA('mle')	SavingsStatusEncoder()
21	0.728	0.662	0.794	RandomForestClassifier()	<NA>	0.740	14.000	1,645.000	5.000	43.000	0.741	entropy	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer(strategy='most_frequent')	None	PCA('mle')	OneHotEncoder()
22	0.721	0.650	0.792	ExtraTreesClassifier()	<NA>	0.781	50.000	590.000	35.000	47.000	0.846	gini	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer(strategy='median')	None	PCA('mle')	SavingsStatusEncoder()
23	0.714	0.698	0.730	XGBClassifier()	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer()	None	None	OneHotEncoder()
24	0.706	0.684	0.728	XGBClassifier()	<NA>	<NA>	3.000	682.000	<NA>	<NA>	<NA>	<NA>	0.152	2.000	0.698	0.940	0.817	0.009	2.086	<NA>	SimpleImputer(strategy='median')	None	None	SavingsStatusEncoder()
25	0.677	0.602	0.752	XGBClassifier()	<NA>	<NA>	15.000	1,159.000	<NA>	<NA>	<NA>	<NA>	0.032	29.000	0.834	0.520	0.503	0.003	1.839	<NA>	SimpleImputer(strategy='median')	None	PCA('mle')	OneHotEncoder()

results.to_formatted_dataframe(query='model == "RandomForestClassifier()"', include_rank=True)

rank	roc_auc Mean	roc_auc 95CI.LO	roc_auc 95CI.HI	max_features	max_depth	n_estimators	min_samples_split	min_samples_leaf	max_samples	criterion	imputer	pca	savings_status_encoder
1	0.763	0.726	0.800	0.114	70.000	553.000	12.000	8.000	0.548	entropy	SimpleImputer()	PCA('mle')	SavingsStatusEncoder()
2	0.758	0.708	0.807	0.681	38.000	1,461.000	23.000	10.000	0.553	gini	SimpleImputer(strategy='median')	None	OneHotEncoder()
3	0.752	0.711	0.792	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer()	None	OneHotEncoder()
4	0.741	0.691	0.791	0.710	15.000	1,493.000	33.000	27.000	0.914	gini	SimpleImputer()	PCA('mle')	OneHotEncoder()
5	0.728	0.662	0.794	0.740	14.000	1,645.000	5.000	43.000	0.741	entropy	SimpleImputer(strategy='most_frequent')	PCA('mle')	OneHotEncoder()

results.to_formatted_dataframe(query='model == "LogisticRegression()"', include_rank=True)

rank	roc_auc Mean	roc_auc 95CI.LO	roc_auc 95CI.HI	C	imputer	scaler	pca	savings_status_encoder
1	0.759	0.713	0.805	<NA>	SimpleImputer()	StandardScaler()	None	OneHotEncoder()
2	0.751	0.682	0.820	0.000	SimpleImputer(strategy='most_frequent')	MinMaxScaler()	None	OneHotEncoder()
3	0.739	0.671	0.807	23.327	SimpleImputer()	StandardScaler()	None	OneHotEncoder()
4	0.733	0.666	0.800	0.000	SimpleImputer(strategy='most_frequent')	StandardScaler()	None	OneHotEncoder()
5	0.729	0.674	0.783	0.001	SimpleImputer(strategy='median')	StandardScaler()	PCA('mle')	SavingsStatusEncoder()

BayesSearchCV Performance Over Time

results.plot_performance_across_trials(facet_by='model').show()

results.plot_performance_across_trials(query='model == "RandomForestClassifier()"').show()

---

Variable Performance Over Time

results.plot_parameter_values_across_trials(query='model == "RandomForestClassifier()"').show()

---

Scatter Matrix

# results.plot_scatter_matrix(query='model == "RandomForestClassifier()"',
#                             height=1000, width=1000).show()

---

Variable Performance - Numeric

results.plot_performance_numeric_params(query='model == "RandomForestClassifier()"',
                                        height=800)

results.plot_parallel_coordinates(query='model == "RandomForestClassifier()"').show()

---

Variable Performance - Non-Numeric

results.plot_performance_non_numeric_params(query='model == "RandomForestClassifier()"').show()

---

results.plot_score_vs_parameter(
    query='model == "RandomForestClassifier()"',
    parameter='max_features',
    size='max_depth',
    color='savings_status_encoder',
)

---

# results.plot_parameter_vs_parameter(
#     query='model == "XGBClassifier()"',
#     parameter_x='colsample_bytree',
#     parameter_y='learning_rate',
#     size='max_depth'
# )

# results.plot_parameter_vs_parameter(
#     query='model == "XGBClassifier()"',
#     parameter_x='colsample_bytree',
#     parameter_y='learning_rate',
#     size='imputer'
# )

---

Last Run - Test Set Performance

last_model = experiment.last_run.download_artifact(
    artifact_name='model/model.pkl',
    read_from=read_pickle,
)
print(type(last_model.model))

<class 'sklearn.pipeline.Pipeline'>

last_model

SklearnModelWrapper(model=Pipeline(steps=[('prep',
                                           ColumnTransformer(transformers=[('numeric',
                                                                            Pipeline(steps=[('imputer',
                                                                                             TransformerChooser(transformer=SimpleImputer())),
                                                                                            ('scaler',
                                                                                             TransformerChooser()),
                                                                                            ('pca',
                                                                                             TransformerChooser(transformer=PCA(n_components='mle')))]),
                                                                            ['duration',
                                                                             'credit_amount',
                                                                             'installment_commitment',
                                                                             'residence_since',
                                                                             'age',
                                                                             'exist...
                                                                            Pipeline(steps=[('savings_encoder',
                                                                                             TransformerChooser(transformer=<source.library.pipeline.SavingsStatusEncoder object at 0xffff0c971950>))]),
                                                                            ['savings_status'])])),
                                          ('model',
                                           RandomForestClassifier(criterion='entropy',
                                                                  max_depth=70,
                                                                  max_features=0.1142268477118407,
                                                                  max_samples=0.5483119512487002,
                                                                  min_samples_leaf=8,
                                                                  min_samples_split=12,
                                                                  n_estimators=553,
                                                                  random_state=42))]))

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

test_predictions = last_model.predict(X_test)
test_predictions[0:10]

array([0.38364602, 0.42461214, 0.50768729, 0.37617143, 0.2083136 ,
       0.33349066, 0.1781312 , 0.4180775 , 0.21162107, 0.25179165])

evaluator = hlp.sklearn_eval.TwoClassEvaluator(
    actual_values=y_test,
    predicted_scores=test_predictions,
    score_threshold=0.37,
)

evaluator.plot_actual_vs_predict_histogram()

evaluator.plot_confusion_matrix()

evaluator.all_metrics_df(return_style=True,
                         dummy_classifier_strategy=['prior', 'constant'],
                         round_by=3)

	Score	Dummy (prior)	Dummy (constant)	Explanation
AUC	0.782	0.500	0.500	Area under the ROC curve (true pos. rate vs false pos. rate); ranges from 0.5 (purely random classifier) to 1.0 (perfect classifier)
True Positive Rate	0.576	0.000	1.000	57.6% of positive instances were correctly identified.; i.e. 34 "Positive Class" labels were correctly identified out of 59 instances; a.k.a Sensitivity/Recall
True Negative Rate	0.844	1.000	0.000	84.4% of negative instances were correctly identified.; i.e. 119 "Negative Class" labels were correctly identified out of 141 instances
False Positive Rate	0.156	0.000	1.000	15.6% of negative instances were incorrectly identified as positive; i.e. 22 "Negative Class" labels were incorrectly identified as "Positive Class", out of 141 instances
False Negative Rate	0.424	1.000	0.000	42.4% of positive instances were incorrectly identified as negative; i.e. 25 "Positive Class" labels were incorrectly identified as "Negative Class", out of 59 instances
Positive Predictive Value	0.607	0.000	0.295	When the model claims an instance is positive, it is correct 60.7% of the time; i.e. out of the 56 times the model predicted "Positive Class", it was correct 34 times; a.k.a precision
Negative Predictive Value	0.826	0.705	0.000	When the model claims an instance is negative, it is correct 82.6% of the time; i.e. out of the 144 times the model predicted "Negative Class", it was correct 119 times
F1 Score	0.591	0.000	0.456	The F1 score can be interpreted as a weighted average of the precision and recall, where an F1 score reaches its best value at 1 and worst score at 0.
Precision/Recall AUC	0.620	0.295	0.295	Precision/Recall AUC is calculated with `average_precision` which summarizes a precision-recall curve as the weighted mean of precisions achieved at each threshold. See sci-kit learn documentation for caveats.
Accuracy	0.765	0.705	0.295	76.5% of instances were correctly identified
Error Rate	0.235	0.295	0.705	23.5% of instances were incorrectly identified
% Positive	0.295	0.295	0.295	29.5% of the data are positive; i.e. out of 200 total observations; 59 are labeled as "Positive Class"
Total Observations	200	200	200	There are 200 total observations; i.e. sample size

evaluator.plot_roc_auc_curve().show()

evaluator.plot_precision_recall_auc_curve().show()

evaluator.plot_threshold_curves(score_threshold_range=(0.1, 0.7)).show()

evaluator.plot_precision_recall_tradeoff(score_threshold_range=(0.1, 0.6)).show()

evaluator.calculate_lift_gain(return_style=True)

/usr/local/lib/python3.11/site-packages/helpsk/sklearn_eval.py:2480: FutureWarning:

The default of observed=False is deprecated and will be changed to True in a future version of pandas. Pass observed=False to retain current behavior or observed=True to adopt the future default and silence this warning.

	Gain	Lift
Percentile
5	0.15	3.05
10	0.20	2.03
15	0.31	2.03
20	0.44	2.20
25	0.53	2.10
30	0.59	1.98
35	0.66	1.89
40	0.71	1.78
45	0.76	1.69
50	0.80	1.59
55	0.80	1.45
60	0.81	1.36
65	0.92	1.41
70	0.93	1.33
75	0.95	1.27
80	0.95	1.19
85	0.97	1.14
90	1.00	1.11
95	1.00	1.05
100	1.00	1.00

---

Production Model - Test Set Performance

production_model = production_run.download_artifact(
    artifact_name='model/model.pkl',
    read_from=read_pickle,
)
print(type(production_model.model))

<class 'sklearn.pipeline.Pipeline'>

production_model

SklearnModelWrapper(model=Pipeline(steps=[('prep',
                                           ColumnTransformer(transformers=[('numeric',
                                                                            Pipeline(steps=[('imputer',
                                                                                             TransformerChooser(transformer=SimpleImputer())),
                                                                                            ('scaler',
                                                                                             TransformerChooser()),
                                                                                            ('pca',
                                                                                             TransformerChooser())]),
                                                                            ['duration',
                                                                             'credit_amount',
                                                                             'installment_commitment',
                                                                             'residence_since',
                                                                             'age',
                                                                             'existing_credits',
                                                                             'num_dependents']),
                                                                           ('n...
                                                                             'employment',
                                                                             'personal_status',
                                                                             'other_parties',
                                                                             'property_magnitude',
                                                                             'other_payment_plans',
                                                                             'housing',
                                                                             'job',
                                                                             'own_telephone',
                                                                             'foreign_worker']),
                                                                           ('savings_status',
                                                                            Pipeline(steps=[('savings_encoder',
                                                                                             TransformerChooser(transformer=OneHotEncoder(handle_unknown='ignore')))]),
                                                                            ['savings_status'])])),
                                          ('model',
                                           RandomForestClassifier(n_estimators=500,
                                                                  random_state=42))]))

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

test_predictions = production_model.predict(X_test)
test_predictions[0:10]

array([0.408, 0.522, 0.678, 0.404, 0.088, 0.454, 0.092, 0.492, 0.176,
       0.232])

evaluator = hlp.sklearn_eval.TwoClassEvaluator(
    actual_values=y_test,
    predicted_scores=test_predictions,
    score_threshold=0.37,
)

evaluator.plot_actual_vs_predict_histogram()

evaluator.plot_confusion_matrix()

evaluator.all_metrics_df(return_style=True,
                         dummy_classifier_strategy=['prior', 'constant'],
                         round_by=3)

	Score	Dummy (prior)	Dummy (constant)	Explanation
AUC	0.815	0.500	0.500	Area under the ROC curve (true pos. rate vs false pos. rate); ranges from 0.5 (purely random classifier) to 1.0 (perfect classifier)
True Positive Rate	0.729	0.000	1.000	72.9% of positive instances were correctly identified.; i.e. 43 "Positive Class" labels were correctly identified out of 59 instances; a.k.a Sensitivity/Recall
True Negative Rate	0.801	1.000	0.000	80.1% of negative instances were correctly identified.; i.e. 113 "Negative Class" labels were correctly identified out of 141 instances
False Positive Rate	0.199	0.000	1.000	19.9% of negative instances were incorrectly identified as positive; i.e. 28 "Negative Class" labels were incorrectly identified as "Positive Class", out of 141 instances
False Negative Rate	0.271	1.000	0.000	27.1% of positive instances were incorrectly identified as negative; i.e. 16 "Positive Class" labels were incorrectly identified as "Negative Class", out of 59 instances
Positive Predictive Value	0.606	0.000	0.295	When the model claims an instance is positive, it is correct 60.6% of the time; i.e. out of the 71 times the model predicted "Positive Class", it was correct 43 times; a.k.a precision
Negative Predictive Value	0.876	0.705	0.000	When the model claims an instance is negative, it is correct 87.6% of the time; i.e. out of the 129 times the model predicted "Negative Class", it was correct 113 times
F1 Score	0.662	0.000	0.456	The F1 score can be interpreted as a weighted average of the precision and recall, where an F1 score reaches its best value at 1 and worst score at 0.
Precision/Recall AUC	0.660	0.295	0.295	Precision/Recall AUC is calculated with `average_precision` which summarizes a precision-recall curve as the weighted mean of precisions achieved at each threshold. See sci-kit learn documentation for caveats.
Accuracy	0.780	0.705	0.295	78.0% of instances were correctly identified
Error Rate	0.220	0.295	0.705	22.0% of instances were incorrectly identified
% Positive	0.295	0.295	0.295	29.5% of the data are positive; i.e. out of 200 total observations; 59 are labeled as "Positive Class"
Total Observations	200	200	200	There are 200 total observations; i.e. sample size

evaluator.plot_roc_auc_curve().show()

evaluator.plot_precision_recall_auc_curve().show()

evaluator.plot_threshold_curves(score_threshold_range=(0.1, 0.7)).show()

evaluator.plot_precision_recall_tradeoff(score_threshold_range=(0.1, 0.6)).show()

evaluator.calculate_lift_gain(return_style=True)

/usr/local/lib/python3.11/site-packages/helpsk/sklearn_eval.py:2480: FutureWarning:

The default of observed=False is deprecated and will be changed to True in a future version of pandas. Pass observed=False to retain current behavior or observed=True to adopt the future default and silence this warning.

	Gain	Lift
Percentile
5	0.14	2.71
10	0.22	2.20
15	0.36	2.37
20	0.51	2.54
25	0.54	2.17
30	0.68	2.26
35	0.73	2.08
40	0.76	1.91
45	0.76	1.69
50	0.81	1.63
55	0.85	1.54
60	0.85	1.41
65	0.88	1.36
70	0.90	1.28
75	0.95	1.27
80	0.97	1.21
85	0.98	1.16
90	0.98	1.09
95	1.00	1.05
100	1.00	1.00

---

Feature Importance

try:
    importances = production_model.model['model'].feature_importances_
    feature_names = [
        x.replace('non_numeric__', '').replace('numeric__', '')
        for x in production_model.model[:-1].get_feature_names_out()
    ]
    feature_importances = sorted(
        zip(feature_names, importances, strict=True),
        key=lambda x: x[1],
        reverse=False,
    )
    fig = px.bar(
        pd.DataFrame(feature_importances, columns=['feature', 'importance']).tail(20),
        y='feature',
        x='importance',
        orientation='h',
        height=700,
        width=800,
        title='Feature Importances of Production Model',
    )
    fig.show()
except:  # noqa
    print("Error calculating feature importances.")

---